Episode 38:47
Nov 19: Banks Say Your AI Agent's Losses Are Yours
TechCrunch reports that Citi, PNC and Citizens now tell customers they are responsible for what outside AI agents do with their accounts, and Citi says that holds even if the agent is hacked. Replit's new Missions and Microsoft's Autopilot both put budgets, approvals and named managers around agents. Plus a family HVAC and plumbing shop whose phone agent first booked emergencies for the next day.
Last updated:
The stories
Citi, PNC and Citizens: outside AI agents are the customer's risk
TechCrunch reports Citi updated its Online User Agreement so customers who give third party AI agents access are responsible, even if the tool is hacked, while Citi's own agents are exempt; PNC and Citizens followed JPMorgan and Bank of America.
Replit Missions: agent swarms with a budget, approvals and a stop rule
Replit's blog says Missions runs groups of agents toward a goal within a spending cap and approval rules you set, with a separate AI judge checking results. Its performance figures are the company's own controlled tests.
Microsoft starts rolling out Autopilot, an always on Copilot agent
Microsoft says Autopilot has its own Entra Agent ID, a named manager and logged actions, and can check with you before key actions. It is starting with Frontier customers, with no price given.
My Bot Did That
Air One Air Conditioning, Heating, Plumbing & Electric, a family business with about 12 employees: Answering after hours and weekend calls and booking service appointments
- What it did
- ServiceTitan's virtual agent, named Jenny, books appointments in English and Spanish and hands calls to a person when needed; after it first offered next day slots for plumbing emergencies, Air One rebuilt its job types around urgency so emergencies can be booked the same night.
- The result, as reported
- By ServiceTitan's customer story, citing Air One's own call tracking: between March and mid June nearly every lead converted to a booked call, against an estimated 30% for the old answering service, figures the article says are not directly comparable. (ServiceTitan (opens in a new tab), )
- Check before you copy it
- Make sure your job types define what an emergency is, ask the vendor about price and who reviews the calls, test the handoff yourself, and send safety calls straight to a qualified person.
Questions and answers
What is this episode about?
TechCrunch reports that Citi, PNC and Citizens now tell customers they are responsible for what outside AI agents do with their accounts, and Citi says that holds even if the agent is hacked. Replit's new Missions and Microsoft's Autopilot both put budgets, approvals and named managers around agents. Plus a family HVAC and plumbing shop whose phone agent first booked emergencies for the next day.
What is the story behind "Citi, PNC and Citizens: outside AI agents are the customer's risk"?
TechCrunch reports Citi updated its Online User Agreement so customers who give third party AI agents access are responsible, even if the tool is hacked, while Citi's own agents are exempt; PNC and Citizens followed JPMorgan and Bank of America. Sources: TechCrunch.
What is the story behind "Replit Missions: agent swarms with a budget, approvals and a stop rule"?
Replit's blog says Missions runs groups of agents toward a goal within a spending cap and approval rules you set, with a separate AI judge checking results. Its performance figures are the company's own controlled tests. Sources: Replit.
What is the story behind "Microsoft starts rolling out Autopilot, an always on Copilot agent"?
Microsoft says Autopilot has its own Entra Agent ID, a named manager and logged actions, and can check with you before key actions. It is starting with Frontier customers, with no price given. Sources: Microsoft Copilot Blog.
What real business use of AI was featured in this episode?
Air One Air Conditioning, Heating, Plumbing & Electric, a family business with about 12 employees: Answering after hours and weekend calls and booking service appointments. ServiceTitan's virtual agent, named Jenny, books appointments in English and Spanish and hands calls to a person when needed; after it first offered next day slots for plumbing emergencies, Air One rebuilt its job types around urgency so emergencies can be booked the same night.
What results were reported, and what should a small firm check first?
By ServiceTitan's customer story, citing Air One's own call tracking: between March and mid June nearly every lead converted to a booked call, against an estimated 30% for the old answering service, figures the article says are not directly comparable. Make sure your job types define what an emergency is, ask the vendor about price and who reviews the calls, test the handoff yourself, and send safety calls straight to a qualified person. Source: ServiceTitan (2026-07-24).
Full transcript
This transcript is the episode script as published, checked against the sources listed above.
Cold Open
Citi, PNC and Citizens have now told customers who hand their bank login to an outside AI agent that what it does is on them. Citi's terms go further: the losses are yours even if the agent gets hacked. That's according to TechCrunch, and it's the first thing to know before you let a bot near your money.
Intro
I'm Jack King, and this is My Bot Did That, the daily show about AI agents doing real work for small firms.
Story 1
TechCrunch reported on Wednesday, November 18, that more big banks are putting the risk of AI agents on the customer. Citi updated its Online User Agreement. According to TechCrunch, if a customer gives a third party AI tool access, and that tool leaks their credentials or gets hacked, the customer is responsible. A Citi spokesperson told TechCrunch that customers who choose to grant account access to third party AI agents are responsible. One wrinkle: Citi's terms say the agents Citi itself provides can be used without that extra liability. So the bank's own bot gets a pass, and everyone else's is your problem.
It isn't only Citi. TechCrunch says JPMorgan and Bank of America announced similar policies last month. PNC, the eighth largest bank, changed its terms the same day the article ran. And Citizens emailed its customers this week with the same kind of warning. TechCrunch also says banks are telling customers they won't be held responsible for reversing a transaction an assistant makes, whether it meant to make it or not.
Why this matters to a small firm. An AI agent, in plain words, is software you give a goal and some access, and it takes steps on its own. A lot of these tools pitch bookkeeping help: read the account, sort the spending, pay the bills. To do that, some ask for your bank login. The TechCrunch piece is about personal finance tools, so if you bank as a business, read your own agreement rather than guessing. But the direction is clear. The bank is not your safety net here.
Say a six person staffing firm wires a bookkeeping agent into its main account. If its bank's terms read like these and that tool gets breached, the bank's position would be that the loss is the firm's.
My take. Treat a bank login like a signed blank check. If a tool wants it, ask whether it can do the job with read only access instead. TechCrunch notes that Anthropic's new personal finance tool is billed as read only, and says it can't move, manage or invest your money. That's the shape you want. And anything that moves money keeps a person on the approve button. Every time.
Story 2
Replit launched something called Missions on November 17. According to the company's blog, you describe the result you want, and a group of AI agents, which Replit calls a swarm, plans, builds, tests and keeps going toward it, within the budget and boundaries you set. The agents run in the cloud, so in Replit's words, you can close your laptop and let them work.
The interesting part for a small firm isn't the swarm. It's the paperwork. Before a mission starts, you agree on the rules. Replit's own example goal is winning your first 10 paying enterprise customers. You can cap the spending at $1,000, and require approval before the agent contacts prospects, shares a proposal, changes customer facing sales materials, or makes product commitments. You choose how often you hear from it: immediately, in a daily digest, or weekly. If the budget runs out first, the work stops with the outcome still unresolved.
There's also what Replit calls an independent AI judge, which is a separate agent that checks the results against the success criteria you agreed on. Replit says that in 2 days, with a $3,000 budget, Missions built a working monitoring platform. That's the company's own test, and Replit itself says these controlled tests don't establish how it performs against the original platform's hosted service. The post doesn't give a price per mission.
My take. Notice that Replit's example is a sales goal, not a coding one. Steal the agreement idea even if you never touch Replit. A spending cap, a short list of things the agent has to ask before doing, and a stop rule. That's more structure than a lot of firms give a new hire on day one. What could go wrong: one AI grading another AI's homework is not a review. If the work touches customers or money, the judge that counts is you. And a spending cap limits what it spends, not what it gets wrong.
Story 3
Microsoft used its Ignite conference on November 17 to start rolling out Autopilot. Microsoft's Copilot blog calls it a trusted autonomous agent that is personal to you, always on, and working for you. Copilot, if you haven't met it, is the AI assistant inside Microsoft 365, the Outlook and Teams world a lot of small offices already live in.
According to Microsoft's post, Autopilot lives in your tenant, which is Microsoft's word for your company's own Microsoft 365 setup, and it has its own identity, memory and tools. Each one gets its own Microsoft Entra Agent ID and a named manager. So the agent has an identity of its own, and the post says its sign ins and actions are logged under its own name.
Microsoft's examples: it can watch Teams channels, follow up on emails, run recurring work, and manage a project. It can even set up and manage a full supplier review process. You decide what work it takes on, and it can check with you before key actions. The catch: it's starting with what Microsoft calls Frontier customers, and the post gives no price for Autopilot.
Why it matters. If your firm already runs on Microsoft 365, this may be the agent that shows up in your inbox first. And the design is sensible. A separate identity with a named human manager means you can see what it did, and there's one thing to switch off if it goes sideways. If you have no IT department, ask whoever looks after your Microsoft account whether you're in Frontier at all. If you aren't, none of this lands on your desk this week, which gives you time to plan.
My take. "Named manager" is the most useful phrase in that post. Pick the person before you turn it on. Decide what counts as a key action at your firm, like sending anything to a client or paying anything, and make it ask first. Then read the follow up emails it sends for a week. An agent that chases people on your behalf can annoy a good customer faster than you can say sorry.
My Bot Did That
Time for the My Bot Did That story.
This one comes from ServiceTitan, the software company whose system the business already used for scheduling, in a customer story dated July 24, 2026. So it's ServiceTitan's account of its own product, and it's from the summer, not today.
The business is Air One Air Conditioning, Heating, Plumbing and Electric, a family business with about 12 employees. The job is after hours and weekend phone calls. Before, those calls went to an answering service that took the caller's information but couldn't book the service call. Air One estimates the old setup converted somewhere around 30 percent of after hours calls.
Air One went live in January with ServiceTitan's virtual agent, which it named Jenny. Jenny books appointments, and it handles both English and Spanish. Setup meant deciding what it could book, and when it needed to hand a call to a person instead.
Here's the honest part. Early on, the agent was offering next day appointments for plumbing emergencies. Air One rebuilt its job types around urgency and gave Jenny access to after hours technician shifts, so emergencies can now be booked the same night. The reported result, from Air One's own call tracking: between March and mid June, nearly every lead converted to a booked call. The article itself says that figure and the 30 percent estimate aren't measuring the same period, season, or quite the same thing.
What to check before you copy it. First, the agent only knows what your job types tell it. If emergency isn't a category, nothing is an emergency. Second, the story gives no price and doesn't say who reviews Jenny's calls, so ask the vendor both. Third, test the handoff yourself at 11 at night. And anything about safety, like a gas smell, goes straight to a qualified person.
Sponsor
My Bot Did That is sponsored by Talus Ops. Talus sets up Grok Bot AI agents in the tools small firms already use, and there's a free guide for your industry at talusops.com/guides.
Close
Today: banks say an outside agent's mistakes are your losses, Replit and Microsoft are building agents that ask before they act, and a shop of about 12 people learned its bot needed to be told what an emergency is. One question to ask at your firm tomorrow: which of our tools holds a password to something that moves money, and who gave it that? If this was useful, subscribe.
AI disclosure
AI disclosure: Jack King is an AI voice, and this episode was written and fact-checked with AI. The audio file carries a C2PA content credential that says so.